-
Notifications
You must be signed in to change notification settings - Fork 13.9k
Optimize checked_ilog
and pow
when base
is a power of two
#147250
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: master
Are you sure you want to change the base?
Optimize checked_ilog
and pow
when base
is a power of two
#147250
Conversation
8f19ce6
to
e5ca8cd
Compare
This comment has been minimized.
This comment has been minimized.
does this affect the codegen in practical circumstances? I would expect even a fairly weak optimizer to perform this optimization, making us hand-coding it irrelevant and potentially even harmful because we introduce more conditional logic to chew through. |
checked_ilog
when base
is a power of twochecked_ilog
and pow
when base
is a power of two
It replaces a loop by some bit manipulations. Whether anyone actually calls either function with a compile-time known, power of 2 base is another question. But it can't hurt, since it is guarded by
No, in either function, LLVM is not able to see that the loop is performing a log/pow and apply the identities: https://godbolt.org/z/6vMsxc9Kh
They're guarded by |
2faa397
to
abb7f32
Compare
...then I'm kinda surprised! Nice catch. |
Would be good to have codegen tests to demonstrate what this is doing -- especially since that way there's a way for people to try removing the special cases later if they think that LLVM no longer needs them. |
abb7f32
to
9ab0f63
Compare
9ab0f63
to
1d0ac82
Compare
thanks for the codegen tests! |
1d0ac82
to
c84b99e
Compare
@rustbot ready |
@scottmcm ping? |
#[no_mangle] | ||
pub fn checked_ilog16(val: u32) -> Option<u32> { | ||
// CHECK: %[[ICMP:.+]] = icmp ne i32 %val, 0 | ||
// CHECK: %[[CTZ:.+]] = tail call range(i32 0, 33) i32 @llvm.ctlz.i32(i32 %val, i1 true) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This isn't really this PR's problem, but consider filing an LLVM bug (assuming it's also true in trunk) that this range is wider than necessary -- we're passing true for is_zero_poison
https://llvm.org/docs/LangRef.html#llvm-ctlz-intrinsic so the range here should be range(i32 0, 32)
.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The range
attribute is too large, but it seems LLVM is still able to propagate the knowledge:
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The code here is looking good, but can you make sure there's normal runtime behaviour tests for it too? Notably, after the base conversation (that's fixed in the code) it made me think that that's not currently covered by any tests, so we should have something -- maybe just add to the # Examples
that it returns None for base zero or one? (They seem like perfectly reasonable and helpful examples, in addition to giving coverage for this stuff.)
And maybe add some should_panic
tests to ensure that the overflow checking is still correct for pow
when overflow checks are enabled?
if base == 2 ** k, then log(base, n) == log(2, n) / k
c84b99e
to
946ce96
Compare
This PR was rebased onto a different master commit. Here's a range-diff highlighting what actually changed. Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers. |
946ce96
to
e481662
Compare
This comment has been minimized.
This comment has been minimized.
e481662
to
264cefa
Compare
This comment has been minimized.
This comment has been minimized.
if base == 2 ** k, then (2 ** k) ** n == 2 ** (k * n) == 1 << (k * n)
264cefa
to
53449c9
Compare
Optimize
checked_ilog
andpow
when the base is a power of two